Approximate-computation-based binary weight convolution neural network hardware accelerator calculating module

Abstract

The invention discloses an approximate-computation-based binary weight convolution neural network hardware accelerator calculating module. The hardware accelerator calculating module is able to receive the input neural element data and binary convolution kernel data and conducts rapid convolution data multiplying, accumulating and calculating. The calculation module utilizes the complement data representation, and includes mainly an optimized approximation binary multiplier, a compressor tree, an innovative approximation adder, and a temporary register for the sum of the serially adding part. In addition, targeted to the optimized approximation binary multiplier, two error compensation schemes are proposed, which reduces or completely eliminates the errors brought about from the optimized approximation binary multiplier under the condition of only slightly increasing the hardware resource overhead expense. Through the optimized calculating units, the key paths for the binary weight convolution neural network hardware accelerator using the computation module are shortened considerably, and the size loss and power loss are also reduced, making the module suitable for a low power consuming embedded type system in need of using the convolution neural network.

Description

technical field [0001] The present invention designs the field of computer and electronic information technology, and in particular relates to a calculation module of a binary weight convolutional neural network hardware accelerator based on approximate calculation. Background technique [0002] The deep convolutional neural network model has made great breakthroughs and successes in many fields such as image classification, action detection, speech recognition and other big data analysis tasks. On the one hand, as the effect of the convolutional neural network becomes better and better, its topology structure is also deepening, and the number of parameters has reached 10 to the 6th power and above, which brings extreme computational complexity. With a big boost, the required computing power has exploded. On the other hand, embedded systems can only provide limited resources, and their power consumption is also limited within a certain range. Although the existing solutions...

AI Technical Summary

Problems solved by technology

[0006] The present invention aims to solve the technical problem of binary weight convolutional neural network applied to embedded systems with limited power consumption and real-time requirements, or at least propose a useful commercial option

Images